PaperlayerView exact revision on arXiv

Source-linked AI summary

PepLLM: ESM-Guided Llama for Structured Protein-Peptide Binding Interface Analysis

Hao Qian, Shikui Tu, Lei Xu

arXiv:2608.21367v1q-bio.BMcs.AIcs.LG

TL;DR

Existing protein–peptide methods provide limited insight into the physicochemical mechanisms underlying binding. PepLLM introduces a dataset-driven ESM–LLaMA framework that generates structured interface annotations, and preliminary results support the feasibility of this mechanism-aware generation task while identifying scope boundaries in the current dataset.

  • Problem

    Existing computational methods mainly address interaction classification, binding-site prediction, or binder generation, leaving limited unified insight into physicochemical binding mechanisms.

  • Method

    PepLLM combines structurally derived JSON supervision with an ESM encoder, nonlinear modality adapter, and LLaMA decoder for instruction-tuned multi-property interface annotation.

  • Results

    PepLLM reliably generates valid JSON, improves most categorical fields and count-field MAE over training, while ESM+adapter remains competitive for some isolated labels.

  • Takeaways & Limitations

    The results support formulating protein–peptide binding interface analysis as structured generation rather than only as independent classification.

  • Takeaways & Limitations

    The dataset excludes complexes with proteins longer than 500 residues, peptides shorter than 2 residues, serialized targets over 2000 tokens, or invalid electrostatic statistics.

Abstract

from arXiv · show

Protein-peptide interactions are central to cellular regulation and peptide-based drug discovery, yet existing computational methods mainly focus on interaction classification, binding-site prediction, or peptide binder generation. These formulations provide limited insight into the physicochemical mechanisms that determine how a peptide binds to a protein. In this work, we introduce \textbf{PepLLM}, an instruction-tuned framework for structured protein-peptide interface understanding. Given protein-peptide sequences, PepLLM generates a machine-readable JSON annotation describing multiple interface properties, including peptide burial state, hydrogen-bond density, salt-bridge presence, hotspot residues, hydrophobicity, and electrostatic complementarity. To support this task, we construct a new protein-peptide interface dataset by integrating structural interface analysis, solvent-accessible surface area computation, hydrophobic burial estimation, electrostatic potential calculation, and redundancy-aware data splitting. PepLLM connects a pretrained ESM encoder with a LLaMA decoder through a nonlinear modality adapter. The adapted ESM residue embeddings are injected into the LLaMA prompt as continuous soft tokens via placeholder-token replacement, enabling the decoder to generate structured interface annotations under instruction tuning. By moving beyond single-label prediction toward multi-property and mechanism-aware generation, PepLLM establishes a new task and modeling paradigm for interpretable protein-peptide interface analysis.

1 Introduction

Protein–peptide interfaces are flexible and governed by multiple physicochemical factors, but existing computational tasks rarely unify these mechanisms. PepLLM addresses this gap with structured JSON generation built on a new annotated dataset and an ESM–LLaMA framework.

  • Protein–peptide binding is difficult to characterize because peptide flexibility, variable poses, and geometric, chemical, and electrostatic factors jointly shape affinity.
  • Existing methods mainly target interaction classification, binding-site prediction, residue-level interface prediction, or peptide binder generation rather than unified mechanism-aware understanding.
  • PepLLM introduces structured protein–peptide interface understanding, generating machine-readable JSON descriptions of burial, hydrogen bonds, salt bridges, hotspots, hydrophobicity, and electrostatic complementarity.
  • The new dataset combines structural interface analysis, solvent-accessible surface area, hydrophobic burial, electrostatic potential, and hotspot extraction with redundancy-aware splitting.
  • PepLLM connects ESM residue representations to a LLaMA decoder through a nonlinear adapter and placeholder-based continuous-token injection for instruction-tuned structured generation.
  • The work moves protein–peptide modeling beyond binary prediction and localization toward interpretable, multi-property, mechanism-aware reasoning.

2 Method

PepLLM combines structural interface annotation with an ESM–LLaMA generation framework to produce structured, mechanism-oriented protein–peptide interface descriptions. The method preserves residue-level ESM representations, injects them into LLaMA through an adapter, and trains generation of JSON annotations.

  • Model framework: PepLLM maps protein–peptide sequence representations to JSON annotations covering burial, hydrogen bonds, salt bridges, hotspots, hydrophobicity, and electrostatic complementarity.The framework targets unified prediction of multiple physicochemical interface properties rather than a single label.
  • Embedding injection: The adapter-aligned ESM embeddings replace reserved LLaMA placeholder-token embeddings, allowing protein representations to function as continuous soft tokens.The ESM attention mask prevents padded hidden states from entering the LLaMA context.
  • Dataset construction: The dataset derives interface labels from structural parsing, solvent-accessible surface areas, hydrophobic burial, electrostatic potentials, and hotspot residues.These heterogeneous structural signals are merged into unified JSON supervision.
  • Interface annotation: The structural pipeline discretizes peptide burial using thresholds of 0.30 and 0.60 and hydrophobicity using thresholds of 0.40 and 0.60.Burial labels are surface, partial, or deep; hydrophobicity labels are polar, mixed, or hydrophobic.
  • Dataset construction: Redundancy-aware splitting clusters protein sequences with MMseqs2 and assigns train, validation, and test sets at the cluster level.The remaining clusters use 0.60, 0.23, and 0.17 ratios with random seed 42.
  • Model framework: PepLLM combines a pretrained ESM encoder, nonlinear modality adapter, and pretrained LLaMA decoder.ESM produces residue-level contextual representations, the adapter projects them into LLaMA’s embedding space, and LLaMA generates the annotation.
  • Model framework: Residue-level ESM hidden states are preserved without pooling for downstream injection into the language model.This design retains positional residue information rather than collapsing the sequence into one representation.
  • Training: Training uses a causal language-modeling objective applied only to target JSON tokens, encouraging valid structured annotations conditioned on instructions and protein–peptide embeddings.Prompt and padding labels are masked with −100, while target-token positions contribute to the loss.

3 Experiments

The experiments evaluate PepLLM for five-field structured interface annotation, valid JSON generation, count-valued predictions, and comparison with an ESM+adapter classification baseline.

  • Experimental Setup: The dataset uses five categorical interface attributes: burial state, electrostatic complementarity, hydrogen-bond density, hydrophobicity, and salt-bridge presence.Examples contain approximately 32K training, 1.8K validation, and 1.8K test instances after sequence clustering and filtering.
  • Models: PepLLM generates interface annotations as JSON, while the ESM+adapter baseline directly predicts five categorical labels with multi-head classification.The comparison separates structured LLM generation from the underlying protein sequence representation.
  • Main Results: 100.00% JSON pass rate at Epoch 1 falls to 99.73% at Epoch 10, indicating consistently valid output formatting throughout training.The JSON pass rate measures whether generated strings can be parsed as structured output.
  • Main Results: Average five-field accuracy increases from 49.14% at Epoch 1 to 52.26% at Epoch 10, with largest gains in hbonds.density and burial state.hbonds.density rises from 42.57% to 48.11%, while burial state rises from 60.93% to 64.40%; electrostatic complementarity remains most challenging.
  • Main Results: Hotspot count MAE decreases from 4.15 to 3.69, whereas salt-bridge count MAE increases slightly from 1.56 to 1.61 across training.The evaluation reports MAE for count-valued JSON fields, where lower values are better.
  • Comparison with ESM-only Classification: The ESM+adapter baseline reaches 54.21% average accuracy versus PepLLM’s 52.26%, while PepLLM performs better on burial state, hbonds.density, and salt bridges.presence.The baseline performs better on hydrophobicity and electrostatic complementarity.
  • Discussion: PepLLM provides unified JSON output containing categorical labels and count-valued interaction information, despite competitive isolated-field performance from discriminative heads.The experiments therefore assess both structured generation and single-field classification.

4 Conclusion

PepLLM combines structured interface parsing with instruction-tuned language-model generation to produce physicochemically grounded descriptions of protein–peptide interactions. The framework points toward interpretable, mechanism-aware molecular interaction analysis and several downstream applications.

  • PepLLM constructs a dataset with structural interface parsing, hydrophobic burial estimation, electrostatic complementarity computation, and cluster-based splitting.
  • The framework generates structured descriptions of multiple interface mechanisms rather than focusing only on interaction classification, localization, contact prediction, or binder generation.
  • ESM residue embeddings are injected into LLaMA as continuous soft tokens through a lightweight modality adapter for valid JSON generation.
  • PepLLM suggests applications in peptide drug screening, docking interpretation, interface engineering, and automated scientific hypothesis generation.

A.1 Overview

The pipeline constructs directed cross-docked peptide–protein pose pairs from clustered complexes. Each pair uses one cluster member as a donor template and another as an acceptor target.

  • The dataset is organized into clusters of peptide–protein complexes for constructing cross-docked poses used in downstream LLM training.
  • Each complex is identified by its PDB identifier, peptide chain, and receptor chain.
  • Within each cluster, every member can provide a donor template for every other member serving as an acceptor target.
  • A directed cross-docking pair assigns the template peptide to the donor and the target receptor to the acceptor.

A.2 Step 1: Standardization and Quality Control

Step 1 standardizes clustered complexes into peptide and receptor structures, computes chain-level quality-control features, and assigns each complex an exportability status.

  • Standardization: Each complex is parsed from an input PDB/mmCIF directory, with peptide and receptor chains extracted using the complex identifier.
  • Standardization: Only the first model is retained when structures contain multiple models, and standardized files are written for the complex, peptide, and receptor.
  • Quality Control: Quality control counts residues missing backbone atoms and consecutive residues whose Cα distance exceeds 4.5 Å for both chains.
  • Quality Control: Complexes receive one of three statuses: ok, repairable, or failed.
  • Quality Control: A complex fails when required structure or chains are unavailable or empty, while repairable denotes exportable chains with missing backbone atoms.
  • Outputs: The step produces complex, pair, and cluster tables together with a quality-control report.

A.3 Step 1.5: Pair Pre-screening and Top-K Re-ranking

Step 1.5 reduces the potentially quadratic directed-pair search by defining receptor sites, cheaply ranking compatible donors, and retaining only top candidates for exact sequence-based re-ranking.

  • Pre-screening: Because the full directed pair table can approach O(n^2), a two-stage ranking procedure is applied before docking.
  • Binding-Site Definition: The receptor binding site comprises residues contacting the native peptide through heavy atoms and is expanded by a one-residue sequence-padding window.
  • Pre-screening: Donor quality is scored from Step 1 QC features, assigning scores of 0 for ok, 0.15 for repairable, and 1.0 for failed complexes.
  • Pre-screening: Prescreening represents peptide and receptor-site sequences as amino-acid composition vectors and requires non-zero peptide and site lengths for valid candidates.
  • Two-Stage Ranking: For each acceptor, only the top M donors by the prescreening score are retained before exact global sequence identities re-rank the candidates.
  • Two-Stage Ranking: The final ranking uses donor–acceptor peptide sequence identity, after which the top K donors are selected for output.
  • Outputs: The procedure outputs a complex feature table, a ranked pair table, and a top-K pair table.

A.4 Step 2: Local Binding-Site Alignment and Initial Model Construction

Step 2 aligns donor and acceptor receptor sequences and binding sites, then transfers the donor peptide into the acceptor receptor frame to construct an initial cross-docked model.

  • Native binding sites are recomputed for donor and acceptor structures using the same heavy-atom contact rule.
  • Donor and acceptor receptor sequences are globally aligned to construct mapped site pairs.
  • If mapped site pairs are insufficient, the pipeline falls back to full-chain mapped Cα pairs; pairs still below the threshold are rejected.
  • A rigid-body superposition transfers the donor peptide into the acceptor receptor coordinate system and defines the initial cross-docked model.
  • The acceptor native peptide is retained only as a reference and is excluded from model construction.

A.5 Step 3: ADCP Local-Docking Job Selection

Step 3 selects successful cross-docking pairs with initial models for ADCP local docking using acceptance criteria and optional heuristic filters.

  • Only successful Step 2 pairs with existing initial models are considered for ADCP local docking.
  • Pairs are accepted according to stated criteria, with an additional extrapolation-ratio filter when that ratio is available.
  • Optional requirements can further filter jobs, including rejection of trivial native-like cases with identical receptor and peptide sequence identities.
  • Accepted jobs are written to the designated job output location.

A.6 Step 4: Preparing PDB Inputs for HPC Docking

Step 4 prepares accepted models and docking inputs for HPC execution, runs ADCP to produce ranked peptide poses, reconstructs final complexes, and organizes structured training records.

  • Input preparation: Initial models are copied into a flat docking-input directory, with large-scale transfers supported through file lists, tar chunks, and remote extraction.
  • Input preparation: For each input complex, the ADCP driver splits receptor and peptide structures, extracts the peptide sequence, protonates both components, and converts them to PDBQT.
  • Local docking: ADCP uses a local docking target around the current peptide pose and outputs a ranked set of docked peptide poses.
  • Complex reconstruction: Docked peptide poses are merged with the corresponding receptor, normalized to peptide chain A and protein chain B, and stored by ranked pose.
  • Training records: Each training example records inputs, targets, and metadata such as local RMSD, site-pair count, peptide length, clash counts, ADCP rank, and source cluster.
  • Dataset organization: Approximately 30,000 directed peptide–protein cross-docking pairs were generated, while cluster-level splitting keeps related pairs within the same dataset split.
Loading 2608.21367v1…